4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/file.h>
23 #include <linux/writeback.h>
24 #include <linux/blkdev.h>
25 #include <linux/buffer_head.h> /* for try_to_release_page(),
26 buffer_heads_over_limit */
27 #include <linux/mm_inline.h>
28 #include <linux/pagevec.h>
29 #include <linux/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/notifier.h>
35 #include <linux/rwsem.h>
37 #include <asm/tlbflush.h>
38 #include <asm/div64.h>
40 #include <linux/swapops.h>
42 /* possible outcome of pageout() */
44 /* failed to write page out, page is locked */
46 /* move page to the active list, page is locked */
48 /* page has been sent to the disk successfully, page is unlocked */
50 /* page is clean and locked */
55 /* Incremented by the number of inactive pages that were scanned */
56 unsigned long nr_scanned
;
58 /* Incremented by the number of pages reclaimed */
59 unsigned long nr_reclaimed
;
61 unsigned long nr_mapped
; /* From page_state */
63 /* This context's GFP mask */
68 /* Can pages be swapped as part of reclaim? */
71 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
72 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
73 * In this context, it doesn't matter that we scan the
74 * whole list at once. */
79 * The list of shrinker callbacks used by to apply pressure to
84 struct list_head list
;
85 int seeks
; /* seeks to recreate an obj */
86 long nr
; /* objs pending delete */
89 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
91 #ifdef ARCH_HAS_PREFETCH
92 #define prefetch_prev_lru_page(_page, _base, _field) \
94 if ((_page)->lru.prev != _base) { \
97 prev = lru_to_page(&(_page->lru)); \
98 prefetch(&prev->_field); \
102 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
105 #ifdef ARCH_HAS_PREFETCHW
106 #define prefetchw_prev_lru_page(_page, _base, _field) \
108 if ((_page)->lru.prev != _base) { \
111 prev = lru_to_page(&(_page->lru)); \
112 prefetchw(&prev->_field); \
116 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
120 * From 0 .. 100. Higher means more swappy.
122 int vm_swappiness
= 60;
123 static long total_memory
;
125 static LIST_HEAD(shrinker_list
);
126 static DECLARE_RWSEM(shrinker_rwsem
);
129 * Add a shrinker callback to be called from the vm
131 struct shrinker
*set_shrinker(int seeks
, shrinker_t theshrinker
)
133 struct shrinker
*shrinker
;
135 shrinker
= kmalloc(sizeof(*shrinker
), GFP_KERNEL
);
137 shrinker
->shrinker
= theshrinker
;
138 shrinker
->seeks
= seeks
;
140 down_write(&shrinker_rwsem
);
141 list_add_tail(&shrinker
->list
, &shrinker_list
);
142 up_write(&shrinker_rwsem
);
146 EXPORT_SYMBOL(set_shrinker
);
151 void remove_shrinker(struct shrinker
*shrinker
)
153 down_write(&shrinker_rwsem
);
154 list_del(&shrinker
->list
);
155 up_write(&shrinker_rwsem
);
158 EXPORT_SYMBOL(remove_shrinker
);
160 #define SHRINK_BATCH 128
162 * Call the shrink functions to age shrinkable caches
164 * Here we assume it costs one seek to replace a lru page and that it also
165 * takes a seek to recreate a cache object. With this in mind we age equal
166 * percentages of the lru and ageable caches. This should balance the seeks
167 * generated by these structures.
169 * If the vm encounted mapped pages on the LRU it increase the pressure on
170 * slab to avoid swapping.
172 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
174 * `lru_pages' represents the number of on-LRU pages in all the zones which
175 * are eligible for the caller's allocation attempt. It is used for balancing
176 * slab reclaim versus page reclaim.
178 * Returns the number of slab objects which we shrunk.
180 unsigned long shrink_slab(unsigned long scanned
, gfp_t gfp_mask
,
181 unsigned long lru_pages
)
183 struct shrinker
*shrinker
;
184 unsigned long ret
= 0;
187 scanned
= SWAP_CLUSTER_MAX
;
189 if (!down_read_trylock(&shrinker_rwsem
))
190 return 1; /* Assume we'll be able to shrink next time */
192 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
193 unsigned long long delta
;
194 unsigned long total_scan
;
195 unsigned long max_pass
= (*shrinker
->shrinker
)(0, gfp_mask
);
197 delta
= (4 * scanned
) / shrinker
->seeks
;
199 do_div(delta
, lru_pages
+ 1);
200 shrinker
->nr
+= delta
;
201 if (shrinker
->nr
< 0) {
202 printk(KERN_ERR
"%s: nr=%ld\n",
203 __FUNCTION__
, shrinker
->nr
);
204 shrinker
->nr
= max_pass
;
208 * Avoid risking looping forever due to too large nr value:
209 * never try to free more than twice the estimate number of
212 if (shrinker
->nr
> max_pass
* 2)
213 shrinker
->nr
= max_pass
* 2;
215 total_scan
= shrinker
->nr
;
218 while (total_scan
>= SHRINK_BATCH
) {
219 long this_scan
= SHRINK_BATCH
;
223 nr_before
= (*shrinker
->shrinker
)(0, gfp_mask
);
224 shrink_ret
= (*shrinker
->shrinker
)(this_scan
, gfp_mask
);
225 if (shrink_ret
== -1)
227 if (shrink_ret
< nr_before
)
228 ret
+= nr_before
- shrink_ret
;
229 mod_page_state(slabs_scanned
, this_scan
);
230 total_scan
-= this_scan
;
235 shrinker
->nr
+= total_scan
;
237 up_read(&shrinker_rwsem
);
241 /* Called without lock on whether page is mapped, so answer is unstable */
242 static inline int page_mapping_inuse(struct page
*page
)
244 struct address_space
*mapping
;
246 /* Page is in somebody's page tables. */
247 if (page_mapped(page
))
250 /* Be more reluctant to reclaim swapcache than pagecache */
251 if (PageSwapCache(page
))
254 mapping
= page_mapping(page
);
258 /* File is mmap'd by somebody? */
259 return mapping_mapped(mapping
);
262 static inline int is_page_cache_freeable(struct page
*page
)
264 return page_count(page
) - !!PagePrivate(page
) == 2;
267 static int may_write_to_queue(struct backing_dev_info
*bdi
)
269 if (current
->flags
& PF_SWAPWRITE
)
271 if (!bdi_write_congested(bdi
))
273 if (bdi
== current
->backing_dev_info
)
279 * We detected a synchronous write error writing a page out. Probably
280 * -ENOSPC. We need to propagate that into the address_space for a subsequent
281 * fsync(), msync() or close().
283 * The tricky part is that after writepage we cannot touch the mapping: nothing
284 * prevents it from being freed up. But we have a ref on the page and once
285 * that page is locked, the mapping is pinned.
287 * We're allowed to run sleeping lock_page() here because we know the caller has
290 static void handle_write_error(struct address_space
*mapping
,
291 struct page
*page
, int error
)
294 if (page_mapping(page
) == mapping
) {
295 if (error
== -ENOSPC
)
296 set_bit(AS_ENOSPC
, &mapping
->flags
);
298 set_bit(AS_EIO
, &mapping
->flags
);
304 * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
306 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
)
309 * If the page is dirty, only perform writeback if that write
310 * will be non-blocking. To prevent this allocation from being
311 * stalled by pagecache activity. But note that there may be
312 * stalls if we need to run get_block(). We could test
313 * PagePrivate for that.
315 * If this process is currently in generic_file_write() against
316 * this page's queue, we can perform writeback even if that
319 * If the page is swapcache, write it back even if that would
320 * block, for some throttling. This happens by accident, because
321 * swap_backing_dev_info is bust: it doesn't reflect the
322 * congestion state of the swapdevs. Easy to fix, if needed.
323 * See swapfile.c:page_queue_congested().
325 if (!is_page_cache_freeable(page
))
329 * Some data journaling orphaned pages can have
330 * page->mapping == NULL while being dirty with clean buffers.
332 if (PagePrivate(page
)) {
333 if (try_to_free_buffers(page
)) {
334 ClearPageDirty(page
);
335 printk("%s: orphaned page\n", __FUNCTION__
);
341 if (mapping
->a_ops
->writepage
== NULL
)
342 return PAGE_ACTIVATE
;
343 if (!may_write_to_queue(mapping
->backing_dev_info
))
346 if (clear_page_dirty_for_io(page
)) {
348 struct writeback_control wbc
= {
349 .sync_mode
= WB_SYNC_NONE
,
350 .nr_to_write
= SWAP_CLUSTER_MAX
,
355 SetPageReclaim(page
);
356 res
= mapping
->a_ops
->writepage(page
, &wbc
);
358 handle_write_error(mapping
, page
, res
);
359 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
360 ClearPageReclaim(page
);
361 return PAGE_ACTIVATE
;
363 if (!PageWriteback(page
)) {
364 /* synchronous write or broken a_ops? */
365 ClearPageReclaim(page
);
374 static int remove_mapping(struct address_space
*mapping
, struct page
*page
)
377 return 0; /* truncate got there first */
379 write_lock_irq(&mapping
->tree_lock
);
382 * The non-racy check for busy page. It is critical to check
383 * PageDirty _after_ making sure that the page is freeable and
384 * not in use by anybody. (pagecache + us == 2)
386 if (unlikely(page_count(page
) != 2))
389 if (unlikely(PageDirty(page
)))
392 if (PageSwapCache(page
)) {
393 swp_entry_t swap
= { .val
= page_private(page
) };
394 __delete_from_swap_cache(page
);
395 write_unlock_irq(&mapping
->tree_lock
);
397 __put_page(page
); /* The pagecache ref */
401 __remove_from_page_cache(page
);
402 write_unlock_irq(&mapping
->tree_lock
);
407 write_unlock_irq(&mapping
->tree_lock
);
412 * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
414 static unsigned long shrink_list(struct list_head
*page_list
,
415 struct scan_control
*sc
)
417 LIST_HEAD(ret_pages
);
418 struct pagevec freed_pvec
;
420 unsigned long reclaimed
= 0;
424 pagevec_init(&freed_pvec
, 1);
425 while (!list_empty(page_list
)) {
426 struct address_space
*mapping
;
433 page
= lru_to_page(page_list
);
434 list_del(&page
->lru
);
436 if (TestSetPageLocked(page
))
439 BUG_ON(PageActive(page
));
443 if (!sc
->may_swap
&& page_mapped(page
))
446 /* Double the slab pressure for mapped and swapcache pages */
447 if (page_mapped(page
) || PageSwapCache(page
))
450 if (PageWriteback(page
))
453 referenced
= page_referenced(page
, 1);
454 /* In active use or really unfreeable? Activate it. */
455 if (referenced
&& page_mapping_inuse(page
))
456 goto activate_locked
;
460 * Anonymous process memory has backing store?
461 * Try to allocate it some swap space here.
463 if (PageAnon(page
) && !PageSwapCache(page
)) {
466 if (!add_to_swap(page
, GFP_ATOMIC
))
467 goto activate_locked
;
469 #endif /* CONFIG_SWAP */
471 mapping
= page_mapping(page
);
472 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
473 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
476 * The page is mapped into the page tables of one or more
477 * processes. Try to unmap it here.
479 if (page_mapped(page
) && mapping
) {
481 * No unmapping if we do not swap
486 switch (try_to_unmap(page
, 0)) {
488 goto activate_locked
;
492 ; /* try to free the page below */
496 if (PageDirty(page
)) {
501 if (!sc
->may_writepage
)
504 /* Page is dirty, try to write it out here */
505 switch(pageout(page
, mapping
)) {
509 goto activate_locked
;
511 if (PageWriteback(page
) || PageDirty(page
))
514 * A synchronous write - probably a ramdisk. Go
515 * ahead and try to reclaim the page.
517 if (TestSetPageLocked(page
))
519 if (PageDirty(page
) || PageWriteback(page
))
521 mapping
= page_mapping(page
);
523 ; /* try to free the page below */
528 * If the page has buffers, try to free the buffer mappings
529 * associated with this page. If we succeed we try to free
532 * We do this even if the page is PageDirty().
533 * try_to_release_page() does not perform I/O, but it is
534 * possible for a page to have PageDirty set, but it is actually
535 * clean (all its buffers are clean). This happens if the
536 * buffers were written out directly, with submit_bh(). ext3
537 * will do this, as well as the blockdev mapping.
538 * try_to_release_page() will discover that cleanness and will
539 * drop the buffers and mark the page clean - it can be freed.
541 * Rarely, pages can have buffers and no ->mapping. These are
542 * the pages which were not successfully invalidated in
543 * truncate_complete_page(). We try to drop those buffers here
544 * and if that worked, and the page is no longer mapped into
545 * process address space (page_count == 1) it can be freed.
546 * Otherwise, leave the page on the LRU so it is swappable.
548 if (PagePrivate(page
)) {
549 if (!try_to_release_page(page
, sc
->gfp_mask
))
550 goto activate_locked
;
551 if (!mapping
&& page_count(page
) == 1)
555 if (!remove_mapping(mapping
, page
))
561 if (!pagevec_add(&freed_pvec
, page
))
562 __pagevec_release_nonlru(&freed_pvec
);
571 list_add(&page
->lru
, &ret_pages
);
572 BUG_ON(PageLRU(page
));
574 list_splice(&ret_pages
, page_list
);
575 if (pagevec_count(&freed_pvec
))
576 __pagevec_release_nonlru(&freed_pvec
);
577 mod_page_state(pgactivate
, pgactivate
);
578 sc
->nr_reclaimed
+= reclaimed
;
582 #ifdef CONFIG_MIGRATION
583 static inline void move_to_lru(struct page
*page
)
585 list_del(&page
->lru
);
586 if (PageActive(page
)) {
588 * lru_cache_add_active checks that
589 * the PG_active bit is off.
591 ClearPageActive(page
);
592 lru_cache_add_active(page
);
600 * Add isolated pages on the list back to the LRU.
602 * returns the number of pages put back.
604 unsigned long putback_lru_pages(struct list_head
*l
)
608 unsigned long count
= 0;
610 list_for_each_entry_safe(page
, page2
, l
, lru
) {
618 * Non migratable page
620 int fail_migrate_page(struct page
*newpage
, struct page
*page
)
624 EXPORT_SYMBOL(fail_migrate_page
);
627 * swapout a single page
628 * page is locked upon entry, unlocked on exit
630 static int swap_page(struct page
*page
)
632 struct address_space
*mapping
= page_mapping(page
);
634 if (page_mapped(page
) && mapping
)
635 if (try_to_unmap(page
, 1) != SWAP_SUCCESS
)
638 if (PageDirty(page
)) {
639 /* Page is dirty, try to write it out here */
640 switch(pageout(page
, mapping
)) {
649 ; /* try to free the page below */
653 if (PagePrivate(page
)) {
654 if (!try_to_release_page(page
, GFP_KERNEL
) ||
655 (!mapping
&& page_count(page
) == 1))
659 if (remove_mapping(mapping
, page
)) {
671 EXPORT_SYMBOL(swap_page
);
674 * Page migration was first developed in the context of the memory hotplug
675 * project. The main authors of the migration code are:
677 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
678 * Hirokazu Takahashi <taka@valinux.co.jp>
679 * Dave Hansen <haveblue@us.ibm.com>
680 * Christoph Lameter <clameter@sgi.com>
684 * Remove references for a page and establish the new page with the correct
685 * basic settings to be able to stop accesses to the page.
687 int migrate_page_remove_references(struct page
*newpage
,
688 struct page
*page
, int nr_refs
)
690 struct address_space
*mapping
= page_mapping(page
);
691 struct page
**radix_pointer
;
694 * Avoid doing any of the following work if the page count
695 * indicates that the page is in use or truncate has removed
698 if (!mapping
|| page_mapcount(page
) + nr_refs
!= page_count(page
))
702 * Establish swap ptes for anonymous pages or destroy pte
705 * In order to reestablish file backed mappings the fault handlers
706 * will take the radix tree_lock which may then be used to stop
707 * processses from accessing this page until the new page is ready.
709 * A process accessing via a swap pte (an anonymous page) will take a
710 * page_lock on the old page which will block the process until the
711 * migration attempt is complete. At that time the PageSwapCache bit
712 * will be examined. If the page was migrated then the PageSwapCache
713 * bit will be clear and the operation to retrieve the page will be
714 * retried which will find the new page in the radix tree. Then a new
715 * direct mapping may be generated based on the radix tree contents.
717 * If the page was not migrated then the PageSwapCache bit
718 * is still set and the operation may continue.
720 if (try_to_unmap(page
, 1) == SWAP_FAIL
)
721 /* A vma has VM_LOCKED set -> Permanent failure */
725 * Give up if we were unable to remove all mappings.
727 if (page_mapcount(page
))
730 write_lock_irq(&mapping
->tree_lock
);
732 radix_pointer
= (struct page
**)radix_tree_lookup_slot(
736 if (!page_mapping(page
) || page_count(page
) != nr_refs
||
737 *radix_pointer
!= page
) {
738 write_unlock_irq(&mapping
->tree_lock
);
743 * Now we know that no one else is looking at the page.
745 * Certain minimal information about a page must be available
746 * in order for other subsystems to properly handle the page if they
747 * find it through the radix tree update before we are finished
751 newpage
->index
= page
->index
;
752 newpage
->mapping
= page
->mapping
;
753 if (PageSwapCache(page
)) {
754 SetPageSwapCache(newpage
);
755 set_page_private(newpage
, page_private(page
));
758 *radix_pointer
= newpage
;
760 write_unlock_irq(&mapping
->tree_lock
);
764 EXPORT_SYMBOL(migrate_page_remove_references
);
767 * Copy the page to its new location
769 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
771 copy_highpage(newpage
, page
);
774 SetPageError(newpage
);
775 if (PageReferenced(page
))
776 SetPageReferenced(newpage
);
777 if (PageUptodate(page
))
778 SetPageUptodate(newpage
);
779 if (PageActive(page
))
780 SetPageActive(newpage
);
781 if (PageChecked(page
))
782 SetPageChecked(newpage
);
783 if (PageMappedToDisk(page
))
784 SetPageMappedToDisk(newpage
);
786 if (PageDirty(page
)) {
787 clear_page_dirty_for_io(page
);
788 set_page_dirty(newpage
);
791 ClearPageSwapCache(page
);
792 ClearPageActive(page
);
793 ClearPagePrivate(page
);
794 set_page_private(page
, 0);
795 page
->mapping
= NULL
;
798 * If any waiters have accumulated on the new page then
801 if (PageWriteback(newpage
))
802 end_page_writeback(newpage
);
804 EXPORT_SYMBOL(migrate_page_copy
);
807 * Common logic to directly migrate a single page suitable for
808 * pages that do not use PagePrivate.
810 * Pages are locked upon entry and exit.
812 int migrate_page(struct page
*newpage
, struct page
*page
)
816 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
818 rc
= migrate_page_remove_references(newpage
, page
, 2);
823 migrate_page_copy(newpage
, page
);
826 * Remove auxiliary swap entries and replace
827 * them with real ptes.
829 * Note that a real pte entry will allow processes that are not
830 * waiting on the page lock to use the new page via the page tables
831 * before the new page is unlocked.
833 remove_from_swap(newpage
);
836 EXPORT_SYMBOL(migrate_page
);
841 * Two lists are passed to this function. The first list
842 * contains the pages isolated from the LRU to be migrated.
843 * The second list contains new pages that the pages isolated
844 * can be moved to. If the second list is NULL then all
845 * pages are swapped out.
847 * The function returns after 10 attempts or if no pages
848 * are movable anymore because to has become empty
849 * or no retryable pages exist anymore.
851 * Return: Number of pages not migrated when "to" ran empty.
853 unsigned long migrate_pages(struct list_head
*from
, struct list_head
*to
,
854 struct list_head
*moved
, struct list_head
*failed
)
857 unsigned long nr_failed
= 0;
861 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
865 current
->flags
|= PF_SWAPWRITE
;
870 list_for_each_entry_safe(page
, page2
, from
, lru
) {
871 struct page
*newpage
= NULL
;
872 struct address_space
*mapping
;
877 if (page_count(page
) == 1)
878 /* page was freed from under us. So we are done. */
881 if (to
&& list_empty(to
))
885 * Skip locked pages during the first two passes to give the
886 * functions holding the lock time to release the page. Later we
887 * use lock_page() to have a higher chance of acquiring the
894 if (TestSetPageLocked(page
))
898 * Only wait on writeback if we have already done a pass where
899 * we we may have triggered writeouts for lots of pages.
902 wait_on_page_writeback(page
);
904 if (PageWriteback(page
))
909 * Anonymous pages must have swap cache references otherwise
910 * the information contained in the page maps cannot be
913 if (PageAnon(page
) && !PageSwapCache(page
)) {
914 if (!add_to_swap(page
, GFP_KERNEL
)) {
921 rc
= swap_page(page
);
925 newpage
= lru_to_page(to
);
929 * Pages are properly locked and writeback is complete.
930 * Try to migrate the page.
932 mapping
= page_mapping(page
);
936 if (mapping
->a_ops
->migratepage
) {
938 * Most pages have a mapping and most filesystems
939 * should provide a migration function. Anonymous
940 * pages are part of swap space which also has its
941 * own migration function. This is the most common
942 * path for page migration.
944 rc
= mapping
->a_ops
->migratepage(newpage
, page
);
949 * Default handling if a filesystem does not provide
950 * a migration function. We can only migrate clean
951 * pages so try to write out any dirty pages first.
953 if (PageDirty(page
)) {
954 switch (pageout(page
, mapping
)) {
960 unlock_page(newpage
);
964 ; /* try to migrate the page below */
969 * Buffers are managed in a filesystem specific way.
970 * We must have no buffers or drop them.
972 if (!page_has_buffers(page
) ||
973 try_to_release_page(page
, GFP_KERNEL
)) {
974 rc
= migrate_page(newpage
, page
);
979 * On early passes with mapped pages simply
980 * retry. There may be a lock held for some
981 * buffers that may go away. Later
986 * Persistently unable to drop buffers..... As a
987 * measure of last resort we fall back to
990 unlock_page(newpage
);
992 rc
= swap_page(page
);
997 unlock_page(newpage
);
1003 if (rc
== -EAGAIN
) {
1006 /* Permanent failure */
1007 list_move(&page
->lru
, failed
);
1011 /* Successful migration. Return page to LRU */
1012 move_to_lru(newpage
);
1014 list_move(&page
->lru
, moved
);
1017 if (retry
&& pass
++ < 10)
1021 current
->flags
&= ~PF_SWAPWRITE
;
1023 return nr_failed
+ retry
;
1027 * Isolate one page from the LRU lists and put it on the
1028 * indicated list with elevated refcount.
1031 * 0 = page not on LRU list
1032 * 1 = page removed from LRU list and added to the specified list.
1034 int isolate_lru_page(struct page
*page
)
1038 if (PageLRU(page
)) {
1039 struct zone
*zone
= page_zone(page
);
1040 spin_lock_irq(&zone
->lru_lock
);
1041 if (PageLRU(page
)) {
1045 if (PageActive(page
))
1046 del_page_from_active_list(zone
, page
);
1048 del_page_from_inactive_list(zone
, page
);
1050 spin_unlock_irq(&zone
->lru_lock
);
1058 * zone->lru_lock is heavily contended. Some of the functions that
1059 * shrink the lists perform better by taking out a batch of pages
1060 * and working on them outside the LRU lock.
1062 * For pagecache intensive workloads, this function is the hottest
1063 * spot in the kernel (apart from copy_*_user functions).
1065 * Appropriate locks must be held before calling this function.
1067 * @nr_to_scan: The number of pages to look through on the list.
1068 * @src: The LRU list to pull pages off.
1069 * @dst: The temp list to put pages on to.
1070 * @scanned: The number of pages that were scanned.
1072 * returns how many pages were moved onto *@dst.
1074 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1075 struct list_head
*src
, struct list_head
*dst
,
1076 unsigned long *scanned
)
1078 unsigned long nr_taken
= 0;
1080 unsigned long scan
= 0;
1082 while (scan
++ < nr_to_scan
&& !list_empty(src
)) {
1083 struct list_head
*target
;
1084 page
= lru_to_page(src
);
1085 prefetchw_prev_lru_page(page
, src
, flags
);
1087 BUG_ON(!PageLRU(page
));
1089 list_del(&page
->lru
);
1091 if (likely(get_page_unless_zero(page
))) {
1093 * Be careful not to clear PageLRU until after we're
1094 * sure the page is not being freed elsewhere -- the
1095 * page release code relies on it.
1100 } /* else it is being freed elsewhere */
1102 list_add(&page
->lru
, target
);
1110 * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
1112 static void shrink_cache(unsigned long max_scan
, struct zone
*zone
,
1113 struct scan_control
*sc
)
1115 LIST_HEAD(page_list
);
1116 struct pagevec pvec
;
1117 unsigned long nr_scanned
= 0;
1119 pagevec_init(&pvec
, 1);
1122 spin_lock_irq(&zone
->lru_lock
);
1125 unsigned long nr_taken
;
1126 unsigned long nr_scan
;
1127 unsigned long nr_freed
;
1129 nr_taken
= isolate_lru_pages(sc
->swap_cluster_max
,
1130 &zone
->inactive_list
,
1131 &page_list
, &nr_scan
);
1132 zone
->nr_inactive
-= nr_taken
;
1133 zone
->pages_scanned
+= nr_scan
;
1134 spin_unlock_irq(&zone
->lru_lock
);
1139 nr_scanned
+= nr_scan
;
1140 nr_freed
= shrink_list(&page_list
, sc
);
1142 local_irq_disable();
1143 if (current_is_kswapd()) {
1144 __mod_page_state_zone(zone
, pgscan_kswapd
, nr_scan
);
1145 __mod_page_state(kswapd_steal
, nr_freed
);
1147 __mod_page_state_zone(zone
, pgscan_direct
, nr_scan
);
1148 __mod_page_state_zone(zone
, pgsteal
, nr_freed
);
1150 spin_lock(&zone
->lru_lock
);
1152 * Put back any unfreeable pages.
1154 while (!list_empty(&page_list
)) {
1155 page
= lru_to_page(&page_list
);
1156 BUG_ON(PageLRU(page
));
1158 list_del(&page
->lru
);
1159 if (PageActive(page
))
1160 add_page_to_active_list(zone
, page
);
1162 add_page_to_inactive_list(zone
, page
);
1163 if (!pagevec_add(&pvec
, page
)) {
1164 spin_unlock_irq(&zone
->lru_lock
);
1165 __pagevec_release(&pvec
);
1166 spin_lock_irq(&zone
->lru_lock
);
1169 } while (nr_scanned
< max_scan
);
1170 spin_unlock_irq(&zone
->lru_lock
);
1172 pagevec_release(&pvec
);
1176 * This moves pages from the active list to the inactive list.
1178 * We move them the other way if the page is referenced by one or more
1179 * processes, from rmap.
1181 * If the pages are mostly unmapped, the processing is fast and it is
1182 * appropriate to hold zone->lru_lock across the whole operation. But if
1183 * the pages are mapped, the processing is slow (page_referenced()) so we
1184 * should drop zone->lru_lock around each page. It's impossible to balance
1185 * this, so instead we remove the pages from the LRU while processing them.
1186 * It is safe to rely on PG_active against the non-LRU pages in here because
1187 * nobody will play with that bit on a non-LRU page.
1189 * The downside is that we have to touch page->_count against each page.
1190 * But we had to alter page->flags anyway.
1193 refill_inactive_zone(unsigned long nr_pages
, struct zone
*zone
,
1194 struct scan_control
*sc
)
1196 unsigned long pgmoved
;
1197 int pgdeactivate
= 0;
1198 unsigned long pgscanned
;
1199 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1200 LIST_HEAD(l_inactive
); /* Pages to go onto the inactive_list */
1201 LIST_HEAD(l_active
); /* Pages to go onto the active_list */
1203 struct pagevec pvec
;
1204 int reclaim_mapped
= 0;
1206 if (unlikely(sc
->may_swap
)) {
1212 * `distress' is a measure of how much trouble we're having
1213 * reclaiming pages. 0 -> no problems. 100 -> great trouble.
1215 distress
= 100 >> zone
->prev_priority
;
1218 * The point of this algorithm is to decide when to start
1219 * reclaiming mapped memory instead of just pagecache. Work out
1223 mapped_ratio
= (sc
->nr_mapped
* 100) / total_memory
;
1226 * Now decide how much we really want to unmap some pages. The
1227 * mapped ratio is downgraded - just because there's a lot of
1228 * mapped memory doesn't necessarily mean that page reclaim
1231 * The distress ratio is important - we don't want to start
1234 * A 100% value of vm_swappiness overrides this algorithm
1237 swap_tendency
= mapped_ratio
/ 2 + distress
+ vm_swappiness
;
1240 * Now use this metric to decide whether to start moving mapped
1241 * memory onto the inactive list.
1243 if (swap_tendency
>= 100)
1248 spin_lock_irq(&zone
->lru_lock
);
1249 pgmoved
= isolate_lru_pages(nr_pages
, &zone
->active_list
,
1250 &l_hold
, &pgscanned
);
1251 zone
->pages_scanned
+= pgscanned
;
1252 zone
->nr_active
-= pgmoved
;
1253 spin_unlock_irq(&zone
->lru_lock
);
1255 while (!list_empty(&l_hold
)) {
1257 page
= lru_to_page(&l_hold
);
1258 list_del(&page
->lru
);
1259 if (page_mapped(page
)) {
1260 if (!reclaim_mapped
||
1261 (total_swap_pages
== 0 && PageAnon(page
)) ||
1262 page_referenced(page
, 0)) {
1263 list_add(&page
->lru
, &l_active
);
1267 list_add(&page
->lru
, &l_inactive
);
1270 pagevec_init(&pvec
, 1);
1272 spin_lock_irq(&zone
->lru_lock
);
1273 while (!list_empty(&l_inactive
)) {
1274 page
= lru_to_page(&l_inactive
);
1275 prefetchw_prev_lru_page(page
, &l_inactive
, flags
);
1276 BUG_ON(PageLRU(page
));
1278 BUG_ON(!PageActive(page
));
1279 ClearPageActive(page
);
1281 list_move(&page
->lru
, &zone
->inactive_list
);
1283 if (!pagevec_add(&pvec
, page
)) {
1284 zone
->nr_inactive
+= pgmoved
;
1285 spin_unlock_irq(&zone
->lru_lock
);
1286 pgdeactivate
+= pgmoved
;
1288 if (buffer_heads_over_limit
)
1289 pagevec_strip(&pvec
);
1290 __pagevec_release(&pvec
);
1291 spin_lock_irq(&zone
->lru_lock
);
1294 zone
->nr_inactive
+= pgmoved
;
1295 pgdeactivate
+= pgmoved
;
1296 if (buffer_heads_over_limit
) {
1297 spin_unlock_irq(&zone
->lru_lock
);
1298 pagevec_strip(&pvec
);
1299 spin_lock_irq(&zone
->lru_lock
);
1303 while (!list_empty(&l_active
)) {
1304 page
= lru_to_page(&l_active
);
1305 prefetchw_prev_lru_page(page
, &l_active
, flags
);
1306 BUG_ON(PageLRU(page
));
1308 BUG_ON(!PageActive(page
));
1309 list_move(&page
->lru
, &zone
->active_list
);
1311 if (!pagevec_add(&pvec
, page
)) {
1312 zone
->nr_active
+= pgmoved
;
1314 spin_unlock_irq(&zone
->lru_lock
);
1315 __pagevec_release(&pvec
);
1316 spin_lock_irq(&zone
->lru_lock
);
1319 zone
->nr_active
+= pgmoved
;
1320 spin_unlock(&zone
->lru_lock
);
1322 __mod_page_state_zone(zone
, pgrefill
, pgscanned
);
1323 __mod_page_state(pgdeactivate
, pgdeactivate
);
1326 pagevec_release(&pvec
);
1330 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1332 static void shrink_zone(int priority
, struct zone
*zone
,
1333 struct scan_control
*sc
)
1335 unsigned long nr_active
;
1336 unsigned long nr_inactive
;
1337 unsigned long nr_to_scan
;
1339 atomic_inc(&zone
->reclaim_in_progress
);
1342 * Add one to `nr_to_scan' just to make sure that the kernel will
1343 * slowly sift through the active list.
1345 zone
->nr_scan_active
+= (zone
->nr_active
>> priority
) + 1;
1346 nr_active
= zone
->nr_scan_active
;
1347 if (nr_active
>= sc
->swap_cluster_max
)
1348 zone
->nr_scan_active
= 0;
1352 zone
->nr_scan_inactive
+= (zone
->nr_inactive
>> priority
) + 1;
1353 nr_inactive
= zone
->nr_scan_inactive
;
1354 if (nr_inactive
>= sc
->swap_cluster_max
)
1355 zone
->nr_scan_inactive
= 0;
1359 while (nr_active
|| nr_inactive
) {
1361 nr_to_scan
= min(nr_active
,
1362 (unsigned long)sc
->swap_cluster_max
);
1363 nr_active
-= nr_to_scan
;
1364 refill_inactive_zone(nr_to_scan
, zone
, sc
);
1368 nr_to_scan
= min(nr_inactive
,
1369 (unsigned long)sc
->swap_cluster_max
);
1370 nr_inactive
-= nr_to_scan
;
1371 shrink_cache(nr_to_scan
, zone
, sc
);
1375 throttle_vm_writeout();
1377 atomic_dec(&zone
->reclaim_in_progress
);
1381 * This is the direct reclaim path, for page-allocating processes. We only
1382 * try to reclaim pages from zones which will satisfy the caller's allocation
1385 * We reclaim from a zone even if that zone is over pages_high. Because:
1386 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1388 * b) The zones may be over pages_high but they must go *over* pages_high to
1389 * satisfy the `incremental min' zone defense algorithm.
1391 * Returns the number of reclaimed pages.
1393 * If a zone is deemed to be full of pinned pages then just give it a light
1394 * scan then give up on it.
1396 static void shrink_caches(int priority
, struct zone
**zones
,
1397 struct scan_control
*sc
)
1401 for (i
= 0; zones
[i
] != NULL
; i
++) {
1402 struct zone
*zone
= zones
[i
];
1404 if (!populated_zone(zone
))
1407 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1410 zone
->temp_priority
= priority
;
1411 if (zone
->prev_priority
> priority
)
1412 zone
->prev_priority
= priority
;
1414 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
1415 continue; /* Let kswapd poll it */
1417 shrink_zone(priority
, zone
, sc
);
1422 * This is the main entry point to direct page reclaim.
1424 * If a full scan of the inactive list fails to free enough memory then we
1425 * are "out of memory" and something needs to be killed.
1427 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1428 * high - the zone may be full of dirty or under-writeback pages, which this
1429 * caller can't do much about. We kick pdflush and take explicit naps in the
1430 * hope that some of these pages can be written. But if the allocating task
1431 * holds filesystem locks which prevent writeout this might not work, and the
1432 * allocation attempt will fail.
1434 unsigned long try_to_free_pages(struct zone
**zones
, gfp_t gfp_mask
)
1438 unsigned long total_scanned
= 0;
1439 unsigned long total_reclaimed
= 0;
1440 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1441 unsigned long lru_pages
= 0;
1443 struct scan_control sc
= {
1444 .gfp_mask
= gfp_mask
,
1445 .may_writepage
= !laptop_mode
,
1446 .swap_cluster_max
= SWAP_CLUSTER_MAX
,
1450 inc_page_state(allocstall
);
1452 for (i
= 0; zones
[i
] != NULL
; i
++) {
1453 struct zone
*zone
= zones
[i
];
1455 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1458 zone
->temp_priority
= DEF_PRIORITY
;
1459 lru_pages
+= zone
->nr_active
+ zone
->nr_inactive
;
1462 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
1463 sc
.nr_mapped
= read_page_state(nr_mapped
);
1465 sc
.nr_reclaimed
= 0;
1467 disable_swap_token();
1468 shrink_caches(priority
, zones
, &sc
);
1469 shrink_slab(sc
.nr_scanned
, gfp_mask
, lru_pages
);
1470 if (reclaim_state
) {
1471 sc
.nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
1472 reclaim_state
->reclaimed_slab
= 0;
1474 total_scanned
+= sc
.nr_scanned
;
1475 total_reclaimed
+= sc
.nr_reclaimed
;
1476 if (total_reclaimed
>= sc
.swap_cluster_max
) {
1482 * Try to write back as many pages as we just scanned. This
1483 * tends to cause slow streaming writers to write data to the
1484 * disk smoothly, at the dirtying rate, which is nice. But
1485 * that's undesirable in laptop mode, where we *want* lumpy
1486 * writeout. So in laptop mode, write out the whole world.
1488 if (total_scanned
> sc
.swap_cluster_max
+
1489 sc
.swap_cluster_max
/ 2) {
1490 wakeup_pdflush(laptop_mode
? 0 : total_scanned
);
1491 sc
.may_writepage
= 1;
1494 /* Take a nap, wait for some writeback to complete */
1495 if (sc
.nr_scanned
&& priority
< DEF_PRIORITY
- 2)
1496 blk_congestion_wait(WRITE
, HZ
/10);
1499 for (i
= 0; zones
[i
] != 0; i
++) {
1500 struct zone
*zone
= zones
[i
];
1502 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1505 zone
->prev_priority
= zone
->temp_priority
;
1511 * For kswapd, balance_pgdat() will work across all this node's zones until
1512 * they are all at pages_high.
1514 * If `nr_pages' is non-zero then it is the number of pages which are to be
1515 * reclaimed, regardless of the zone occupancies. This is a software suspend
1518 * Returns the number of pages which were actually freed.
1520 * There is special handling here for zones which are full of pinned pages.
1521 * This can happen if the pages are all mlocked, or if they are all used by
1522 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1523 * What we do is to detect the case where all pages in the zone have been
1524 * scanned twice and there has been zero successful reclaim. Mark the zone as
1525 * dead and from now on, only perform a short scan. Basically we're polling
1526 * the zone for when the problem goes away.
1528 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1529 * zones which have free_pages > pages_high, but once a zone is found to have
1530 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1531 * of the number of free pages in the lower zones. This interoperates with
1532 * the page allocator fallback scheme to ensure that aging of pages is balanced
1535 static unsigned long balance_pgdat(pg_data_t
*pgdat
, unsigned long nr_pages
,
1538 unsigned long to_free
= nr_pages
;
1542 unsigned long total_scanned
;
1543 unsigned long total_reclaimed
;
1544 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1545 struct scan_control sc
= {
1546 .gfp_mask
= GFP_KERNEL
,
1548 .swap_cluster_max
= nr_pages
? nr_pages
: SWAP_CLUSTER_MAX
,
1553 total_reclaimed
= 0;
1554 sc
.may_writepage
= !laptop_mode
,
1555 sc
.nr_mapped
= read_page_state(nr_mapped
);
1557 inc_page_state(pageoutrun
);
1559 for (i
= 0; i
< pgdat
->nr_zones
; i
++) {
1560 struct zone
*zone
= pgdat
->node_zones
+ i
;
1562 zone
->temp_priority
= DEF_PRIORITY
;
1565 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
1566 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
1567 unsigned long lru_pages
= 0;
1569 /* The swap token gets in the way of swapout... */
1571 disable_swap_token();
1575 if (nr_pages
== 0) {
1577 * Scan in the highmem->dma direction for the highest
1578 * zone which needs scanning
1580 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
1581 struct zone
*zone
= pgdat
->node_zones
+ i
;
1583 if (!populated_zone(zone
))
1586 if (zone
->all_unreclaimable
&&
1587 priority
!= DEF_PRIORITY
)
1590 if (!zone_watermark_ok(zone
, order
,
1591 zone
->pages_high
, 0, 0)) {
1598 end_zone
= pgdat
->nr_zones
- 1;
1601 for (i
= 0; i
<= end_zone
; i
++) {
1602 struct zone
*zone
= pgdat
->node_zones
+ i
;
1604 lru_pages
+= zone
->nr_active
+ zone
->nr_inactive
;
1608 * Now scan the zone in the dma->highmem direction, stopping
1609 * at the last zone which needs scanning.
1611 * We do this because the page allocator works in the opposite
1612 * direction. This prevents the page allocator from allocating
1613 * pages behind kswapd's direction of progress, which would
1614 * cause too much scanning of the lower zones.
1616 for (i
= 0; i
<= end_zone
; i
++) {
1617 struct zone
*zone
= pgdat
->node_zones
+ i
;
1620 if (!populated_zone(zone
))
1623 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
1626 if (nr_pages
== 0) { /* Not software suspend */
1627 if (!zone_watermark_ok(zone
, order
,
1628 zone
->pages_high
, end_zone
, 0))
1631 zone
->temp_priority
= priority
;
1632 if (zone
->prev_priority
> priority
)
1633 zone
->prev_priority
= priority
;
1635 sc
.nr_reclaimed
= 0;
1636 shrink_zone(priority
, zone
, &sc
);
1637 reclaim_state
->reclaimed_slab
= 0;
1638 nr_slab
= shrink_slab(sc
.nr_scanned
, GFP_KERNEL
,
1640 sc
.nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
1641 total_reclaimed
+= sc
.nr_reclaimed
;
1642 total_scanned
+= sc
.nr_scanned
;
1643 if (zone
->all_unreclaimable
)
1645 if (nr_slab
== 0 && zone
->pages_scanned
>=
1646 (zone
->nr_active
+ zone
->nr_inactive
) * 4)
1647 zone
->all_unreclaimable
= 1;
1649 * If we've done a decent amount of scanning and
1650 * the reclaim ratio is low, start doing writepage
1651 * even in laptop mode
1653 if (total_scanned
> SWAP_CLUSTER_MAX
* 2 &&
1654 total_scanned
> total_reclaimed
+total_reclaimed
/2)
1655 sc
.may_writepage
= 1;
1657 if (nr_pages
&& to_free
> total_reclaimed
)
1658 continue; /* swsusp: need to do more work */
1660 break; /* kswapd: all done */
1662 * OK, kswapd is getting into trouble. Take a nap, then take
1663 * another pass across the zones.
1665 if (total_scanned
&& priority
< DEF_PRIORITY
- 2)
1666 blk_congestion_wait(WRITE
, HZ
/10);
1669 * We do this so kswapd doesn't build up large priorities for
1670 * example when it is freeing in parallel with allocators. It
1671 * matches the direct reclaim path behaviour in terms of impact
1672 * on zone->*_priority.
1674 if ((total_reclaimed
>= SWAP_CLUSTER_MAX
) && (!nr_pages
))
1678 for (i
= 0; i
< pgdat
->nr_zones
; i
++) {
1679 struct zone
*zone
= pgdat
->node_zones
+ i
;
1681 zone
->prev_priority
= zone
->temp_priority
;
1683 if (!all_zones_ok
) {
1688 return total_reclaimed
;
1692 * The background pageout daemon, started as a kernel thread
1693 * from the init process.
1695 * This basically trickles out pages so that we have _some_
1696 * free memory available even if there is no other activity
1697 * that frees anything up. This is needed for things like routing
1698 * etc, where we otherwise might have all activity going on in
1699 * asynchronous contexts that cannot page things out.
1701 * If there are applications that are active memory-allocators
1702 * (most normal use), this basically shouldn't matter.
1704 static int kswapd(void *p
)
1706 unsigned long order
;
1707 pg_data_t
*pgdat
= (pg_data_t
*)p
;
1708 struct task_struct
*tsk
= current
;
1710 struct reclaim_state reclaim_state
= {
1711 .reclaimed_slab
= 0,
1715 daemonize("kswapd%d", pgdat
->node_id
);
1716 cpumask
= node_to_cpumask(pgdat
->node_id
);
1717 if (!cpus_empty(cpumask
))
1718 set_cpus_allowed(tsk
, cpumask
);
1719 current
->reclaim_state
= &reclaim_state
;
1722 * Tell the memory management that we're a "memory allocator",
1723 * and that if we need more memory we should get access to it
1724 * regardless (see "__alloc_pages()"). "kswapd" should
1725 * never get caught in the normal page freeing logic.
1727 * (Kswapd normally doesn't need memory anyway, but sometimes
1728 * you need a small amount of memory in order to be able to
1729 * page out something else, and this flag essentially protects
1730 * us from recursively trying to free more memory as we're
1731 * trying to free the first piece of memory in the first place).
1733 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
1737 unsigned long new_order
;
1741 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
1742 new_order
= pgdat
->kswapd_max_order
;
1743 pgdat
->kswapd_max_order
= 0;
1744 if (order
< new_order
) {
1746 * Don't sleep if someone wants a larger 'order'
1752 order
= pgdat
->kswapd_max_order
;
1754 finish_wait(&pgdat
->kswapd_wait
, &wait
);
1756 balance_pgdat(pgdat
, 0, order
);
1762 * A zone is low on free memory, so wake its kswapd task to service it.
1764 void wakeup_kswapd(struct zone
*zone
, int order
)
1768 if (!populated_zone(zone
))
1771 pgdat
= zone
->zone_pgdat
;
1772 if (zone_watermark_ok(zone
, order
, zone
->pages_low
, 0, 0))
1774 if (pgdat
->kswapd_max_order
< order
)
1775 pgdat
->kswapd_max_order
= order
;
1776 if (!cpuset_zone_allowed(zone
, __GFP_HARDWALL
))
1778 if (!waitqueue_active(&pgdat
->kswapd_wait
))
1780 wake_up_interruptible(&pgdat
->kswapd_wait
);
1785 * Try to free `nr_pages' of memory, system-wide. Returns the number of freed
1788 unsigned long shrink_all_memory(unsigned long nr_pages
)
1791 unsigned long nr_to_free
= nr_pages
;
1792 unsigned long ret
= 0;
1793 struct reclaim_state reclaim_state
= {
1794 .reclaimed_slab
= 0,
1797 current
->reclaim_state
= &reclaim_state
;
1798 for_each_pgdat(pgdat
) {
1799 unsigned long freed
;
1801 freed
= balance_pgdat(pgdat
, nr_to_free
, 0);
1803 nr_to_free
-= freed
;
1804 if ((long)nr_to_free
<= 0)
1807 current
->reclaim_state
= NULL
;
1812 #ifdef CONFIG_HOTPLUG_CPU
1813 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1814 not required for correctness. So if the last cpu in a node goes
1815 away, we get changed to run anywhere: as the first one comes back,
1816 restore their cpu bindings. */
1817 static int __devinit
cpu_callback(struct notifier_block
*nfb
,
1818 unsigned long action
, void *hcpu
)
1823 if (action
== CPU_ONLINE
) {
1824 for_each_pgdat(pgdat
) {
1825 mask
= node_to_cpumask(pgdat
->node_id
);
1826 if (any_online_cpu(mask
) != NR_CPUS
)
1827 /* One of our CPUs online: restore mask */
1828 set_cpus_allowed(pgdat
->kswapd
, mask
);
1833 #endif /* CONFIG_HOTPLUG_CPU */
1835 static int __init
kswapd_init(void)
1840 for_each_pgdat(pgdat
) {
1843 pid
= kernel_thread(kswapd
, pgdat
, CLONE_KERNEL
);
1845 pgdat
->kswapd
= find_task_by_pid(pid
);
1847 total_memory
= nr_free_pagecache_pages();
1848 hotcpu_notifier(cpu_callback
, 0);
1852 module_init(kswapd_init
)
1858 * If non-zero call zone_reclaim when the number of free pages falls below
1861 * In the future we may add flags to the mode. However, the page allocator
1862 * should only have to check that zone_reclaim_mode != 0 before calling
1865 int zone_reclaim_mode __read_mostly
;
1867 #define RECLAIM_OFF 0
1868 #define RECLAIM_ZONE (1<<0) /* Run shrink_cache on the zone */
1869 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
1870 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
1871 #define RECLAIM_SLAB (1<<3) /* Do a global slab shrink if the zone is out of memory */
1874 * Mininum time between zone reclaim scans
1876 int zone_reclaim_interval __read_mostly
= 30*HZ
;
1879 * Priority for ZONE_RECLAIM. This determines the fraction of pages
1880 * of a node considered for each zone_reclaim. 4 scans 1/16th of
1883 #define ZONE_RECLAIM_PRIORITY 4
1886 * Try to free up some pages from this zone through reclaim.
1888 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
1890 const unsigned long nr_pages
= 1 << order
;
1891 struct task_struct
*p
= current
;
1892 struct reclaim_state reclaim_state
;
1894 struct scan_control sc
= {
1895 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
1896 .may_swap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
1897 .nr_mapped
= read_page_state(nr_mapped
),
1898 .swap_cluster_max
= max_t(unsigned long, nr_pages
,
1900 .gfp_mask
= gfp_mask
,
1903 disable_swap_token();
1906 * We need to be able to allocate from the reserves for RECLAIM_SWAP
1907 * and we also need to be able to write out pages for RECLAIM_WRITE
1910 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
1911 reclaim_state
.reclaimed_slab
= 0;
1912 p
->reclaim_state
= &reclaim_state
;
1915 * Free memory by calling shrink zone with increasing priorities
1916 * until we have enough memory freed.
1918 priority
= ZONE_RECLAIM_PRIORITY
;
1920 shrink_zone(priority
, zone
, &sc
);
1922 } while (priority
>= 0 && sc
.nr_reclaimed
< nr_pages
);
1924 if (sc
.nr_reclaimed
< nr_pages
&& (zone_reclaim_mode
& RECLAIM_SLAB
)) {
1926 * shrink_slab does not currently allow us to determine
1927 * how many pages were freed in the zone. So we just
1928 * shake the slab and then go offnode for a single allocation.
1930 * shrink_slab will free memory on all zones and may take
1933 shrink_slab(sc
.nr_scanned
, gfp_mask
, order
);
1936 p
->reclaim_state
= NULL
;
1937 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
1939 if (sc
.nr_reclaimed
== 0)
1940 zone
->last_unsuccessful_zone_reclaim
= jiffies
;
1942 return sc
.nr_reclaimed
>= nr_pages
;
1945 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
1951 * Do not reclaim if there was a recent unsuccessful attempt at zone
1952 * reclaim. In that case we let allocations go off node for the
1953 * zone_reclaim_interval. Otherwise we would scan for each off-node
1956 if (time_before(jiffies
,
1957 zone
->last_unsuccessful_zone_reclaim
+ zone_reclaim_interval
))
1961 * Avoid concurrent zone reclaims, do not reclaim in a zone that does
1962 * not have reclaimable pages and if we should not delay the allocation
1965 if (!(gfp_mask
& __GFP_WAIT
) ||
1966 zone
->all_unreclaimable
||
1967 atomic_read(&zone
->reclaim_in_progress
) > 0 ||
1968 (current
->flags
& PF_MEMALLOC
))
1972 * Only run zone reclaim on the local zone or on zones that do not
1973 * have associated processors. This will favor the local processor
1974 * over remote processors and spread off node memory allocations
1975 * as wide as possible.
1977 node_id
= zone
->zone_pgdat
->node_id
;
1978 mask
= node_to_cpumask(node_id
);
1979 if (!cpus_empty(mask
) && node_id
!= numa_node_id())
1981 return __zone_reclaim(zone
, gfp_mask
, order
);